Forecasting weekly reference evapotranspiration using Auto Encoder Decoder Bidirectional LSTM model hybridized with a Boruta-CatBoost input optimizer

Abstract

Reference evapotranspiration (ETo) is one of the most important and influential components in optimizing agricultural water consumption and water resources management. In the present study, an efficient deep learning model, Auto Encode Decoder Bidirectional Long Short-Term Memory (AED-BiLSTM), was applied for the first time in forecasting 1–3 weeks ahead of weekly ETo. Three meteorological stations, including Kermanshah (semi-arid), Nowshahr (very humid), and Yazd (arid), which have different climates, were investigated and evaluated. For this purpose, a statistical period of 20 years (2000 to 2019), of which 15 years (2000–2014) for training and the final five years (2015–2015) for model testing were considered. The generalized regression neural network (GRNN) and extreme gradient boosting (XGBoost) machine learning models were used to compare the performance of the newly developed model. A novel CatBoost-Boruta algorithm was used to determine the critical lags in the forecasting process. The results showed that the newly developed model (AED-BiLSTM) in all three meteorological stations had higher ability and accuracy in forecasting weekly ETo than the GRNN and XGBoost models. The outcomes of testing phase in Kermanshah station ascertained that the AED-BiLSTM in terms of (ETo(t + 1): R (correlation coefficient) = 0.9835, RMSE (root mean square error, mm/week) = 3.4597; ETo(t + 2): R = 0.9805, RMSE = 3.7215; ETo(t + 3): R = 0.9795, RMSE = 3.8857) outperformed the other models. In Nowshahr station AED-BiLSTM model had higher accuracy in terms of (ETo(t + 1): R = 0.9057, RMSE = 4.6786; ETo(t + 2): R = 0.9036, RMSE = 4.7697; ETo(t + 3): R = 0.9024, RMSE = 4.8437). Also, in Yazd station AED-BiLSTM perform better than other models in terms of (ETo(t + 1): R = 0.9827, RMSE = 3.1319; ETo(t + 2): R = 0.9800, RMSE = 3.3998; ETo(t + 3): R = 0.9774, RMSE = 3.5679). The XGBoost and GRNN models performed better after the AED-BiLSTM model, respectively. The effect of climate type on the accuracy of the models showed that the models have higher efficiency and accuracy in arid and semi-arid climates.